DECODE: Diagnosing Interpretability Failures in Automated Interpretation Process
4.40T1 sourcearXiv cs.HC
Source record
Published by arXiv cs.HC (T1 source). The original is at https://arxiv.org/abs/2611.17856.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryDECODE is a visual analytics system for diagnosing failures in LLM-assisted interpretability. It separates diagnosis into three stages, uses consensus among independent explanations as a reliability proxy, and recommends cases for manual review through active seeding. A case study examines 16,384 sparse autoencoder features.
Why it mattersThe paper turns a difficult validation problem into explicit review checkpoints, useful to teams designing oversight for AI-assisted analysis.
Cited by
No citations on record.